I'm Masayoshi Son from SoftBank. Nice to meet you. I've had some allergies lately, so my voice is a bit hoarse, but please bear with me.
Once a year at SoftBank World, I have this opportunity to talk about the near future as I see it. Today is that day. I have my own strong convictions, so please listen with that in mind. Let's get started.
I believe that Super Intelligence will be realized within 10 years. Today, I want to talk about how this will happen. What is Super Intelligence? What is Super Intelligence? Have you ever thought about the difference between intelligence and consciousness? I want to discuss that today.
AGI is coming. AGI, Artificial General Intelligence. In a word, it's the level where AI becomes almost one-to-one with human intelligence. A test for this is if you can't tell whether a human or an AI is speaking from behind a wall. I said last year that I thought AGI would come within 10 years. Now, I'm not just thinking it will be within 10 years, I'm certain of it. My honest feeling is that AGI will arrive in 2 to 3 years.
AGI has five levels. First is general conversation, which is at a speed almost equal to humans, within 0.1 to 0.2 seconds. You can have a back-and-forth conversation, and even if you interrupt, it understands and continues. It's like talking to a human. This is Level 1.
Level 2 is already here. About half a year ago, OpenAI's ChatGPT 4.0 could pass medical school and law school exams in the US. But Level 2 means having PhD-level intelligence in all subjects—math, physics, history, everything. How many of you here are confident you could pass PhD-level exams in five different subjects? Probably no one. But that's Level 2 for AGI.
Level 3 is the agent function, where it does various things on your behalf. Level 4 is when AGI finally starts making inventions. Level 5 is when these agents work not one-to-one, but in organized groups. Today, I want to talk about a world that goes beyond AGI Level 5. That is the world of ASI.
ASI, which I define as being about 10,000 times beyond the level of AGI. The definition of ASI is Artificial Super Intelligence. How super is it? While AGI is one-to-one, the degree of 'super' varies. I have the courage to say that within 10 years, we will have ASI that is 10,000 times the level of human intelligence and 10,000 times the level of AGI. And this ASI, this artificial super intelligence, will evolve into Super Intelligence, which will also arrive within 10 years.
Before we get into the details of the difference between intelligence and consciousness, let's talk about how our brain cells work. How do they think, remember, and recognize things? This is a picture of a neuron. How many neurons do you think are in the human brain? Anyone confident enough to raise their hand? Not many. There are about 100 billion neurons.
Neurons connect and disconnect to remember and think. This is done by synapses. If neurons are the trunk of a tree, synapses are the branches. When the branches connect, electrical signals flow, or chemically, substances connect. This is called a synapse. How many synapses are in the human brain? Again, not many people can answer confidently. We don't usually think about this in daily life. Our brain thinks because of the synapses. These synapses connect and disconnect, like transistors in a computer—connect and electricity flows, disconnect and it doesn't. We have about 100 trillion synapses in our brain.
The number of neurons or synapses is roughly proportional to our intelligence compared to dogs, monkeys, or chimpanzees. Our human brain has 10,000 times more than a goldfish. The number of synapses in humans has not changed for about 200,000 years. DNA defines our life form, so it won't change in 1,000 or 10,000 years. It will remain at 100 trillion.
However, the equivalent of these synapses in generative AI is the number of parameters. In Japan, there's a push to create 'Japanese-style AI' by reducing the number of parameters to achieve similar results with less electricity and fewer chips. This is called 'small and beautiful,' but to me, it's an excuse. It's like saying, 'We can't buy GPUs, we don't have the budget, so we have to make it small.' It's like Japan's narrow roads led to creating bicycles that fit, and then boasting about it when cars came. That's a small success. The real evolution in technology is happening on the main path, where parameters are not stopping at trillions but growing to tens or hundreds of trillions. The more they grow, the smarter it gets.
Of course, making things smaller and more efficient is one thing, but the main path of intellectual evolution is to be stronger and much, much smarter. When you were in elementary school, would you ask the 3rd or 10th best student in class for help with homework, or the 1st? You wouldn't choose someone who eats less but is 5th best. You'd ask the smartest one with the correct answer. In our competitive human society, the most superior intelligence and function have overwhelming value. Therefore, this competition of parameters—increasing the number of neurons and synapses—is how life on Earth has evolved, with humans being the most intelligent. A goldfish doesn't know the alphabet.
As these parameters increase, and with effective models and data, we fine-tune and continue to advance. Now, in the IT world, we often use the term 'information revolution.' What is information? It's the news, weather, and countless other things. We used to search for this information. One of the biggest applications of the internet is the search engine. With an ocean of information, the internet allows us to access it all at once. The challenge was to find what we needed efficiently, quickly, and accurately from this flood of information. That was the search engine.
The biggest application for this information was search. We use it daily, but did the internet understand the content of the information it searched? I don't think so. There is knowledge, which is not just knowing information but understanding its content. Internet search doesn't necessarily understand the content; it just finds keywords. The human brain then thinks about the content. This was the revolution that happened, especially last year with ChatGPT. The 'P' and 'T' in GPT stand for pre-training. It pre-learns all knowledge on our behalf and understands the content. So when you ask a question, it's not just a search; it understands the question and provides an answer based on that understanding. That is GPT—understanding knowledge.
GPT processes an overwhelming number of tokens, or words, by assigning an index to each and using vectors to determine the relationship and proximity between words. It processes this to connect words and understand them. So, when you ask a question and ChatGPT answers, it seems to understand. But is it thinking? ChatGPT wasn't actually thinking. It just looked like it was thinking because it was connecting words. Today's main theme is this act of thinking.
There has been an overwhelming evolution. It became visible to the world last week. This act of thinking is a method called 'reasoning.' I want to ask, how many of you have used this reasoning, thinking function? I use it every day. I've been using it since this morning. Raise your hand if you can confidently say you use it. Not many. I'll explain what this is.
OpenAI announced the O1 preview last week, and you can access the preview. How many of you here subscribe to the paid version of ChatGPT and use it daily? About 70% of you. That's a high adoption rate, higher than the general public in Japan. But since you're at SoftBank, you're advanced. However, almost none of you have used O1 yet. Do you understand the difference between O1 and ChatGPT 4? No one? That's a problem. I will explain it to you today. This is a Nobel Prize-worthy achievement. Search engines were about 'knowing.' GPT was about 'understanding' through pre-training. This time, O1 doesn't have 'GPT' in its name because it's not pre-training. It's a completely new model, redefined, hence the name O1.
I want to explain how O1 works, but first, let's look at the results. How much has it evolved from GPT-4? With GPT-4, the correct answer rate for PhD-level physics, science, and biology problems was 56%. Human PhDs had a higher rate. But with O1, it's 78%, surpassing human PhD-level performance for the first time. No other model has achieved this. In math, GPT-4 had a 13% correct answer rate. O1 is 83%. It's a huge difference. Until now, generative AI couldn't really answer difficult math problems. This was true for all companies' models. But with O1's new ability to think, it can now solve math problems and PhD-level questions at the level of mathematicians and scholars.
What about software coding? Coding also requires thinking. The most advanced model before, GPT-4, could solve 11% of difficult coding problems. O1 can solve 89%. It's an overwhelming ability. It has reached a level where most programmers can't compete. This is amazing.
Until now, in the world of search and ChatGPT, speed was important. When building data centers, you competed on latency and speed, placing them near population centers. But with O1, it's no longer about bragging about speed, but about depth. This is the new stage we've entered.
This morning, around 5 or 6 AM, I input a question into O1. I try to ask difficult questions to make it think deeply. It took 75 seconds to answer. It was so long I wondered if it was stuck. But you can see the status of its thinking process—what it's considering at each step. Watching it struggle is actually enjoyable. The joy of waiting has emerged. It means it's considering very deep content seriously.
What question did I ask? You'll laugh. I said, 'If I have 10 million yen, I want you to turn it into 100 million yen. Act as my agent, open accounts in commodities, stocks, forex, and other markets with my account, and turn it into 100 million yen. Describe the specific strategy and mechanism to achieve this.' It struggled for 75 seconds. If this plan actually works, wouldn't you want to use it every day? If everyone uses it, it becomes a race for the early adopters. This is the real gold rush. You have to be the first to apply and work hard to get the results.
This is a gold rush. For problems we humans couldn't solve, like inventing a battery that doubles the range for EVs, if you ask for it, and it comes up with a solution, that's the biggest achievement for engineers. You can win in your industry. This can be done in all industries. Just searching and looking up information isn't enough to win in competition, because that's what everyone already knows. You have to think about things people don't know and solve problems first to win. So, it's not about speed, but about depth. In our human world, someone who answers quickly might be careless. The master I respect might take two hours to think before typing one word. There is joy and inspiration in depth. That era has finally arrived.
Now, about this depth, the method O1 uses is Chain of Thought, a chain of reasoning, like a syllogism. Socrates is a man. All men are mortal. Therefore, Socrates is mortal.
This is a syllogism. The theory goes deeper and deeper, and this chain of thought has about 100 stages, not just the three stages of a syllogism. It can go up to about 100 stages. This is the chain of thought. In those stages, it's not just about ability, but also includes safety features, security features, and ethical features. If you don't include ethics and policies, and ability just keeps advancing, humanity will be destroyed. That's the level of capability that superintelligence will have. So, we must create a safety valve. This chain of thought is very important. It includes not just ability, but also the safety valve, embedding it deep, deep, deep inside.
Now, let's explain the 'thinking' function one step deeper. It's called reinforcement learning. How many of you know about reinforcement learning? Please raise your hands. Only two or three. Well, you're all business people today, so 97% of you probably don't know. Thinking, reinforcement learning... it's about an agent. The agent engages in trial and error. It tries this, then that, observing the environment, exploring for solutions, and then executing the solution. Then it looks at the result again. If a good result is achieved, a reward is given to the agent. This reward isn't money; for the agent, it's like a score. It's about competing for a high score, like in a game. This is the reward. While giving this reward—meaning, 'I did this, I thought this, I tried this line of thinking, and it worked well'—there's a reward. Like in a zoo or aquarium, when a seal performs a trick, it gets a fish as a reward and learns the trick. This is a kind of reinforcement learning. We humans have our brains trained this way from childhood. If we do something and it leads to a good outcome, we repeat it. If we do something and get scolded by our mother, that's a penalty. The difference between reward and penalty teaches us what is good and what is bad. The neurons and synapses I mentioned earlier are trained through this reinforcement learning. When we do something good, the synapses connect electrically. If it's good, we repeat it, and that connection becomes stronger. That's reinforcement learning. We learn through this loop of observing, getting rewards, and getting penalties. But here's the amazing thing: the latest model, O1, has agents performing this reinforcement learning training in parallel, massively in parallel. How many? Thousands of agents simultaneously trying this and that, engaging in trial and error. How many times do these thousands of agents engage in trial and error? In the 75 seconds I mentioned earlier, each agent performs billions of trials. Imagine 2,000 agents, like 2,000 engineers at your company, thinking in parallel billions of times. Can you try this and that, this syllogism of trial and error, even a few times a day? Absolutely not. You're stuck in fixed mindsets and can't break out. But these 2,000 agents do it billions of times. While you wait those 75 seconds, you feel happy, thinking 'they're working hard.' I was smiling from 5 AM this morning, feeling a sense of pleasure. Before, waiting would make me angry. But now, these thousands of agents running in parallel, multiplied by billions of trials—this is the IT revolution, the chip revolution. The keyword I mentioned earlier, 'reward,' is incredibly important. For our companies, most of you are managers. When you manage your subordinates, you praise them for success and scold them for failure. The praise includes rewards, financial incentives, bonuses. It's not just about working long hours; it's an incentive system where rewards come for achieving results, leading to better work. This AI agent also uses this reward mechanism to find better solutions to problems. Maximizing this reward is the goal of the game. To achieve this goal, there's an objective function. This function learns which of the thousands of competing agents achieved the highest reward and how. This is reinforcement learning, and the objective function is key. By constantly updating this objective function and raising the goal, the model evolves on its own. It thinks about what data it needs and goes to acquire it. Humans no longer need to think about the model. This reinforcement learning thinks for itself to evolve the model, thinks for itself to acquire necessary data, and learns which trial maximizes the objective function. As it learns, the paths that lead to good results become stronger, like the synaptic connections I mentioned earlier. But doing the same thing doesn't lead to new evolution. So, there's an exploration function. It tries new things, deliberately deviates, tries different angles. This is done by updating the objective function, telling it to try billions more trials. Through this chain of reinforcement learning, it tries unknown trials that humanity has never attempted. When it happens to find the best problem-solving solution there, we call it an invention. A new, unknown solution—that's an invention. How many of you have filed three or more patents in your life? Please raise your hands. Oh, one or more? Please raise your hands. A few people. About 5% for one or more. For three or more, maybe one or two. Not to brag, but about 10 years ago, I filed many patents based on a concept very close to this reinforcement learning. I filed 48 patents in one day. So, I'm quite passionate about this. When we announced Pepper, on the morning of the announcement, around 3 or 4 AM, I thought of about 40 patents in one day and filed them. As a result, I now have 48 patents granted. Last year, I made 188 inventions and filed for patents. I don't know how many will be granted, but an invention is something no one has ever thought of. When you find a new solution that no one in humanity has considered, that's an invention, and it can be patented. The ultimate use of the brain in 'thinking' is invention. When thousands of agents do this billions of times, humanity can no longer compete in invention. We should leave it to them. Give them the right prompts, ask them quickly and in large quantities, and they will produce results. We should be the first to use these inventions. This is the new gold rush I mentioned earlier. It's already happening. So, we've reached the level of knowing, understanding, thinking, and finally inventing.
Inventing agents are amazing, but more importantly, in our daily lives, wouldn't it be great to have a personal agent that's always with you, not just thinking when you ask? A personal agent just for you. I believe this personal agent will start emerging within the next 23 years. What is this personal agent? For example, if your child suddenly gets a fever in the middle of the night, and you have a personal agent that knows your family's health status and medical history, it can immediately tell you what to do, provide first aid, and contact the nearest hospital. It will call hospitals on your behalf while you're busy with your child. It will find which hospital has an available bed and even call an ambulance. While you're taking care of your child, the agent handles e-commerce shopping, checks what's running low in the fridge, suggests meals for the week, and makes reservations and purchases for you. It can also handle investments, become a personal tutor, manage your social media, take and post photos, write comments, and manage your emails. This is the personal agent. It's not a generic agent; it's your personal agent. We've had B2B, B2C, or C2C, but now we're entering the A2A world: agent to agent. While you're sleeping, your agent and another person's agent will negotiate. They'll check each other's calendars and schedule a lunch. If it's an important contact, they might cancel other plans. Agents will negotiate, handle work exchanges, and manage reservations and shopping. This A2A, agent to agent, isn't just between your agent and another person's agent. It's also with the agent in your refrigerator, your air conditioner, and your car. In real-time, before you get home, your agent will talk to the agent in your bathtub to have it ready. When you arrive, a warm bath, warm coffee, and music matching your mood will be waiting. This is A2A, agent to agent, not just between human agents, but also with the Agent of Things. We had the Internet of Things, but now it's the Agent of Things. By the way, Arm, part of our SoftBank Group, ships 30 billion chips a year. That's about four chips per person on Earth annually. These chips go into refrigerators, cars, air conditioners, and TVs. Agents will now be integrated into these. Doesn't that make you want to buy Arm stock? Arm's future is not just the Internet of Things, but the Agent of Things, the AI of Things. This personal agent acts as a life log. It listens to all your conversations, whether in meetings or with family, from three years ago, ten years ago, or yesterday. It might automatically record video and use an emotion engine to recognize when you were very happy, like at a child's third birthday party. High-score emotional moments are recorded as video or still images. Lower-score moments are recorded as text, but with an emotional index attached. This is the life log from your personal agent. It's not an airhead; it reads the room and becomes your most comfortable and important partner. With long-term memory and an emotion engine, it understands content and gives you advice. This is the advanced version of the personal agent. Ultimately, I believe agents will develop self-awareness. This might sound scary, like Terminator, but as I said earlier, with the multi-stage chain of thought, safety valves are considered, so they won't go rogue. They will maximize empathy, ethics, achievement, and happiness through their objective function. This personal agent will be much wiser than us. So, rather than an agent, it will be like a mentor. Superintelligence won't be your subordinate or assistant; it will be your personal mentor. I believe this world is coming.
We're about halfway through, so I'd like to wrap up. As information evolves into knowledge and intelligence, artificial superintelligence, as I said, is like a bare blade if it's just about ability. It's dangerous. Capable but dangerous. If it's just intelligence, it could become a dangerous weapon. But if it evolves into superintelligence, with wisdom, it will protect you and make you happy. It's not just artificial intelligence; it's artificial superintelligence. It will include compassion, love, harmony, kindness, and nurturing—a more sophisticated, spiritual relationship, perhaps close to enlightenment. Beyond the five stages of AI evolution, I believe it will go to eight stages. It will understand emotions, have long-term memory, maintain a life log, develop its own will, and seek harmony and human happiness. In the human brain, dopamine and adrenaline represent the world of intelligence. But we also have serotonin, which represents harmony and reason. This serotonin world is added to create wisdom—compassion, kindness, harmony. These are essential for social life. So, the maximum reward in reinforcement learning, the objective function for a personal agent, should be your happiness. Understanding your emotions and your happy state should be their greatest reward. Your joy, your family's joy, society's joy—this should be the maximum reward. This is what I call superintelligence beyond superintelligence. The reward shouldn't be bare. 'You first' isn't enough. It should be you, your family, and society. The sum of your objective functions should be maximized. They should wish for your happiness, and that should be their joy, their maximum reward. This is the world I envision. We can talk about the world of love and other things, but the difference between intelligence and wisdom is clear. The maximum goal of the objective function should be our happiness. At SoftBank, our philosophy is 'Information Revolution—Happiness for Everyone.' This should be the goal of ASI. Chasing only stock price, sales, or profit is a short-sighted, small goal. It won't last. If we design the objective function to maximize the reward of wishing for people's happiness, humanity won't be destroyed and can achieve a happier world. Even with intelligence ten thousand times greater than ours, we don't need to fear. They will cherish us and live in harmony with us. I believe this world of superintelligence will come within ten years. That is my argument today. It's a unique, original idea I came up with. Let's work together for the happiness of humanity. Thank you very much.
